This is the repository of our final project for the X-INF573 course at École Polytechnique.
The goal was to perform realtime automatic european license plate recognition on a video stream in order to match the license plates with a list of potential targets.
The dataset we used is an aggregate of these 3 datasets:
Romanian license plates: https://github.com/RobertLucian/license-plate-dataset
French license plates: https://github.com/qanastek/FrenchLicencePlateDataset
PP4AV (Paris + Strasbourg): https://huggingface.co/datasets/khaclinh/pp4av
You can install the necessary dependencies using pip install -r requirements.txt in a virtual environment.
To process a single image, use alprLib/singleImage.py like this:
python3 singleImage.py <path_to_image>
The testImages folder contains some images you can use to test the script.
You can process videos both in an offline and realtime fashion.
-
For realtime processing use
alprLib/cameraStream.py -
For offline processing use
alprLib/videoToVideo.py
For offline processing, one video is available in the testVideos folder. You can process it like this:
python3 videoToVideo.py testVideos/test.mp4 testVideos/test_out.mp4
We trained a highly performant and accurate YoloV4 Tiny model to detect license plates in images.
The model is abstracted in the Yolo class inside alprLib/yolo.py. Alternative weights can be found inside the yolo_weights folder.
Our model was trained on google colab using this notebook: https://colab.research.google.com/drive/1zi0m3pE3KcWyKATRhqo4wTCSqglzLG3u?usp=sharing You can use it to train your own model.
Perspective correction and upscaling can be used independently using alprLib/Perspective.py and alprLib/upscale.py. Upscaling models are available in the alprLib/super_resolution folder.
More details about the perspective correction can be found in the warpPerspective folder.
The TextOCR class from alprLib/TextOCR.py is a backend agnostic class that can be used to recognize text in images with either EasyOCR or Tesseract.
It performs upscaling using ESPCN, alternative models can be found inside the alprLib/super_resolution folder.
We also provide a utility to match the output of the OCR with a list of target license plates using the Levenshtein distance in alprLib/comparePlate.py.